GPU implementation of algorithm SIMPLE-TS for calculation of unsteady, viscous, compressible and heat-conductive gas flows

TL;DR

This paper presents GPU implementations of the SIMPLE-TS algorithm for simulating unsteady, viscous, compressible, heat-conductive gas flows, achieving significant speedups and memory efficiency over CPU counterparts.

Contribution

It introduces optimized GPU algorithms for SIMPLE-TS, demonstrating high performance and simplicity in code structure for complex fluid flow simulations.

Findings

Speedup of up to 184x on AMD Radeon R9 280X compared to CPU

Memory requirement reduced by half compared to CPU implementation

Code is simple, portable, and highly efficient

Abstract

The recent trend of using Graphics Processing Units (GPU's) for high performance computations is driven by the high ratio of price performance for these units, complemented by their cost effectiveness. At first glance, computational fluid dynamics (CFD) solvers match perfectly to GPU resources because these solvers make intensive calculations and use relatively little memory. Nevertheless, there are scarce results about the practical use of this serious advantage of GPU over CPU, especially for calculations of viscous, compressible, heat-conductive gas flows with double precision accuracy. In this paper, two GPU algorithms according to time approximation of convective terms were presented: explicit and implicit scheme. To decrease data transfers between device memories and increase the arithmetic intensity of a GPU code we minimize the number of kernels. The GPU algorithm was implemented in one kernel for the implicit scheme and two kernels for the explicit scheme. The numerical equations were put together using macros and optimization, data copy from global to private memory, and data reuse were left to the compiler. Thus keeps the code simpler with excellent maintenance. As a test case, we model the flow past squares in a microchannel at supersonic speed. The tests show that overall speedup of AMD Radeon R9 280X is up to 102x compared to Intel Core i5-4690 core and up to 184x compared to Intel Core i7-920 core, while speedup of NVIDIA Tesla M2090 is up to 11x compared to Intel Core i5-4690 core and up to 20x compared to Intel Core i7-920 core. Memory requirements of GPU code are improved compared to CPU one. It requires 1[GB] global memory for 5.9 million finite volumes that are two times less compared to C++ CPU code. After all the code is simple, portable (written in OpenCL), memory efficient and easily modifiable moreover demonstrates excellent performance.